Driver behavior recognition method for automatic driving virtual test and verification

Abstract

The invention discloses a driver behavior identification method and system for automatic driving virtual test and verification. Traffic video data are collected through a multi-source sensor, dynamic target object classification and trajectory extraction are performed by using a deep neural network, and a trajectory is converted into global coordinates and stored; and based on the trajectory data, using a deep learning algorithm to analyze driving behavior characteristics in a complex traffic scene, calibrating driver model parameters by considering regional differences, and generating a driver model. And constructing a virtual test framework based on GiA software and PTV Vissim joint simulation, and carrying out virtual test and verification. The driver behavior can be accurately identified, the authenticity and simulation accuracy of a test scene are improved, the model is more practical by considering regional difference, the test efficiency and range are improved by cloud parallel test, and the method has important significance on development and commercialization of an automatic driving system and can also be applied to other fields of intelligent traffic.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving, and particularly relates to a driver behavior recognition method for virtual testing and verification of autonomous driving. Background Art

[0002] With the development of automotive technology towards low carbon dioxide emissions and highly autonomous driving, highly autonomous driving has become a key factor in the intelligent transportation of smart cities, and is expected to realize new transportation concepts such as multimodal transportation, robot taxi services, and intelligent logistics, which is of great significance for energy conservation, emission reduction, and improving traffic efficiency. However, the main obstacle to its commercialization is safety verification, which requires proving that the autonomous driving system is superior to human drivers in vehicle control. Currently, due to time and cost limitations, it is difficult to complete hundreds of millions of kilometers of tests on actual roads to generate sufficient traffic conflict situations through actual road tests. Although the scenario-based testing method can reduce the actual road test mileage, there are problems such as non-representative test scenarios, incomplete test coverage, and limited interaction between the preset trajectory and the driving environment. Traffic flow simulation (TFS) is an alternative, but its simulation authenticity depends to a large extent on the driver model. Existing driver models have serious deficiencies in modeling human lane-changing behavior in complex urban traffic, lack data that can describe human driving behavior in complex urban traffic, and it is difficult to accurately simulate complex lane-changing behaviors, such as lane-changing situations at highway entrances and exits with multiple lanes and densely populated urban intersections. Although traffic flow simulation tools such as PTV Vissim, Aimsun, and SUMO are widely used, due to the randomness of driver and vehicle behavior, it is difficult to analyze traffic scenarios analytically. Model calibration requires a large amount of empirical data and involves numerous parameters, and it is difficult to achieve effective calibration using traditional experimental and error-finding methods. Currently, the verification and calibration of path planning are still in their infancy, and it is difficult to obtain trajectory data of all vehicles, which limits the accuracy of the driver behavior model and the authenticity of the simulation. Summary of the Invention

[0003] The purpose of the present invention is to provide a driver behavior recognition method for virtual testing and verification of autonomous driving. The present invention can improve the authenticity of test scenarios, accurately simulate driver behavior in complex traffic scenarios, and provide more reliable support for the safety verification of autonomous driving systems.

[0004] The technical solution of the present invention: A driver behavior recognition method for virtual testing and verification of autonomous driving includes the following steps:

[0005] S1. Collect traffic video data of the target area through multi-source sensors, and the multi-source sensors include road cameras and GNSS devices;

[0006] S2. Classify the dynamic objects in the traffic video data based on a deep neural network, and extract the movement trajectories of the objects;

[0007] S3. Convert the extracted movement trajectories into trajectory data in the global coordinate system, and associate and store them with GNSS timestamps;

[0008] S4. Based on the trajectory data, use a deep learning algorithm to analyze the driving behavior characteristics in complex traffic scenarios;

[0009] S5. Calibrate the driver model parameters in the traffic flow simulation according to the driving behavior characteristics to generate a driver model;

[0010] S6. Build a virtual test framework and verify the performance of the autonomous driving function under the calibrated driver model through co-simulation.

[0011] In the above method for identifying driver behavior for virtual testing and verification of autonomous driving, in step S2, the dynamic object classification includes passenger cars, trucks, motorcycles, bicycles, and pedestrians, and the continuous extraction of trajectories is achieved through a target detection algorithm and a multi-object tracking algorithm.

[0012] In the aforementioned method for identifying driver behavior for virtual testing and verification of autonomous driving, in step S3, the specific process of converting the image coordinates into global three-dimensional coordinates through a camera calibration model and associating and storing them with GNSS timestamps is as follows:

[0013] Conversion from the image coordinate system to the camera coordinate system:

[0014]

[0015] In the formula: (u, υ) are the image pixel coordinates, Z is the z-axis coordinate of the target point in the camera coordinate system; the camera internal parameter matrix K is obtained through the Zhang Zhengyou calibration method:

[0016]

[0017] In the formula: f x , f y respectively represent the focal length coordinates (in pixel units), (u0, υ0) are the image principal point coordinates,

[0018] Conversion from the camera coordinate system to the world three-dimensional coordinate system:

[0019] P w = R T ·(P c - T);

[0020] In the formula: P w is the three-dimensional point in the world three-dimensional coordinate system, P cis a three-dimensional point in the camera coordinate system; R is a rotation matrix R (3×3 orthogonal matrix); T is a translation vector (3×1 vector); R T is the inverse of the rotation matrix;

[0021] The GNSS timestamp associated storage includes hardware synchronization and time synchronization. The hardware synchronization is ensured through the NTP protocol or hardware trigger signal, and the clock deviation between all cameras and the GNSS receiver is <1ms. The video frame capture time is accurately recorded as tcam; the time synchronization is to store the three-dimensional coordinate P w and the corresponding timestamp tcam as structured data:

[0022]

[0023] The above-mentioned driver behavior recognition method for autonomous driving virtual testing and verification. In step S4, the complex traffic scenarios include highway entrances, exits, and multi-lane intersections, and the driving behavior characteristics include lane-changing decisions, lane-keeping, and following behavior;

[0024] The above-mentioned driver behavior recognition method for autonomous driving virtual testing and verification. The lane-changing decision acquisition includes the following steps:

[0025] S4.1 Calculate the rate of change of the distance between the vehicle and the adjacent lane markings, and set a threshold to determine the starting point of lane change;

[0026] S4.2 Analyze the relative speed and distance of surrounding vehicles based on the Wiedemann psychophysical model to judge the type of lane-changing intention.

[0027] The above-mentioned driver behavior recognition method for autonomous driving virtual testing and verification. In step S5, the regional differences are used to calibrate the model parameters by comparing the driving behavior data sets in different regions, including the aggressiveness of lane change, safety distance, and path planning preferences.

[0028] The above-mentioned driver behavior recognition method for autonomous driving virtual testing and verification. In step S6, the virtual testing framework is implemented based on the joint simulation of GaiA software and PTV Vissim, including static scene modeling, dynamic traffic flow generation, and multi-body vehicle dynamics model integration.

[0029] The above-mentioned driver behavior recognition method for autonomous driving virtual testing and verification. The joint simulation realizes the data interaction between GaiA and Vissim through dynamic link libraries, and supports the closed-loop testing of traffic flow simulation and autonomous driving control algorithms.

[0030] Compared with the prior art, the present invention can accurately identify the driver's behavior through technical means such as multi-source sensor data collection, deep neural network analysis, and joint simulation, providing a reliable method and system for virtual testing and verification of autonomous driving, and improving the authenticity of the test scenario and the accuracy of simulation. The present invention takes into account the influence of regional differences on driving behavior, calibrates the driver model parameters by comparing the driving behavior data sets in different regions, making the model more in line with the actual situation, and helping to more comprehensively evaluate the performance of the autonomous driving system in different regions. The virtual test framework developed by the present invention supports cloud parallel virtual testing, realizes the automated evaluation of a large number of scenarios through distributed computing, greatly improves the test efficiency and coverage, can more quickly discover the problems existing in the autonomous driving system, and accelerates its commercialization process. The present invention can also be applied to other fields of intelligent transportation systems, such as traffic flow prediction, road planning optimization, etc., and has broad application prospects and social and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Shows a schematic diagram of the trajectory tracking effect of the present invention.

[0032] Figure 2 Shows an example diagram of converting image coordinates to global three-dimensional coordinates.

[0033] Figure 3 Shows a schematic diagram of the distribution of trajectory data according to the Wiedemann model.

[0034] Figure 4 Is a randomly generated driving behavior and traffic flow simulation effect. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] The present invention will be further described below in conjunction with embodiments, but it shall not be used as a basis for limiting the present invention.

[0036] Embodiment: A method for identifying driver behavior for virtual testing and verification of autonomous driving includes the following steps:

[0037] S1. Collect traffic video data of the target area through multi-source sensors, and the multi-source sensors include road cameras and GNSS devices; in this step, multiple test sites are selected in Austria and Zhejiang Province, such as highway entrances / exits, multi-lane intersections, etc. Install road cameras and GNSS devices at these sites to ensure that the cameras can clearly capture the traffic scene, and the GNSS devices can accurately record the time stamp and geographical location information. Continuously collect traffic video data and store it at a certain time interval to provide sufficient data resources for subsequent analysis. Figure 1 Shows a traffic video data diagram.

[0038] S2. Classify the dynamic objects in the traffic video data based on a deep neural network, and extract the motion trajectories of the objects; in this step, the collected traffic video data is input into an object recognition and trajectory extraction module based on a deep neural network. The deep neural network is trained with a large amount of traffic image data and can accurately identify dynamic objects such as passenger cars, trucks, motorcycles, bicycles, and pedestrians. The object detection algorithm determines the position of the object in the image, and the multi-object tracking algorithm continuously extracts the trajectory according to the motion characteristics of the object. In this process, the algorithm parameters are continuously optimized to improve the accuracy of object recognition and trajectory extraction.

[0039] In this step, the YOLOv5 model is used to implement multi-class object detection, and its core formulas include:

[0040] 1. Feature extraction stage:

[0041] F = Darknet53(I);

[0042] Where: I is the input image, and F is the extracted multi-scale feature map.

[0043] 2. Bounding box regression loss function:

[0044]

[0045] Where: b i and are the predicted box and the ground truth box respectively, λ coord is the weight coefficient, S 2 represents the number of grids into which the input image is divided; B is the number of bounding boxes predicted for each grid; is an indicator function. If the j-th bounding box in the i-th grid is responsible for detecting an object, w i and h i are the width and height of the i-th bounding box respectively;

[0046] 3. Classification loss function:

[0047]

[0048] Where: p i (c) represents the true probability that the object in the i-th grid belongs to class c, represents the predicted probability that the object in the i-th grid belongs to class c.

[0049] The multi-object tracking algorithm adopts a JPDA-based framework. When using Kalman filter prediction, the state transition equation is:

[0050] x k = Fx k-1 + Bu k-1+w k-1 ;

[0051] Where: x has the same state, F is the state transition matrix, H is the observation matrix; k represents the time, u represents the control input, and w represents the noise;

[0052] The observation equation is:

[0053] z k = Hx k + υ k ;

[0054] Where: υ k represents the observation noise at time k;

[0055] Match using the Hungarian algorithm, cost matrix element:

[0056] C i，j = 1 - cosine(f i , g j ) + α·d i，j ;

[0057] Where: f i and g j are the ReID (person re-identification) feature vectors of the i-th detection box and the j-th trajectory respectively, used to describe the appearance features of the target. Cosine represents the cosine similarity, and d i，j represents the Euclidean distance between the i-th detection box and the j-th trajectory, used to measure their proximity in spatial position, and α is the weight coefficient;

[0058] Trajectory management includes initialization, update, and deletion. If the number of unmatched frames exceeds the threshold θ 2, the trajectory is terminated.

[0059] In terms of spatio-temporal constraints, there are speed constraints υ min ≤ ||Δp i，j || / Δt ≤ υ max and angle constraint Δφ ≤ φ max , where Δp i，j represents the position change vector of the target between adjacent frames, Δt is the time interval between adjacent frames, and Δφ is the angle change of the target between adjacent frames.

[0060] Introduce the ReID model to extract appearance features, feature similarity:

[0061]

[0062] This method combines detection and tracking, and in the highway scenario, MOTA reaches 92.3% and IDF1 reaches 89.7%. Figure 1 Shows the schematic diagram of the trajectory tracking effect of the present invention.

[0063] S3. Convert the extracted motion trajectory into trajectory data in the global coordinate system and store it associated with the GNSS timestamp; in this step, use the camera calibration model to convert the image coordinates into global coordinates, fuse the data from multiple camera perspectives to improve the trajectory accuracy. Associate the converted trajectory data with the GNSS timestamp and store it in the Research Data Management (RDM) system. During the storage process, follow the FAIR principle to ensure the findability, accessibility, interoperability, and reusability of the data, facilitating subsequent data analysis and sharing.

[0064] The specific process of converting the image coordinates into global three-dimensional coordinates through the camera calibration model and storing them associated with the GNSS timestamp is as follows:

[0065] Convert from the image coordinate system to the camera coordinate system:

[0066]

[0067] In the formula: (u, υ) are the image pixel coordinates, and Z is the z-axis coordinate of the target point in the camera coordinate system; obtain the camera internal parameter matrix K through the Zhang Zhengyou calibration method:

[0068]

[0069] In the formula: f x , f y respectively represent the focal length coordinates (in pixel units), and (u0, υ0) are the image principal point coordinates.

[0070] Convert from the camera coordinate system to the world three-dimensional coordinate system:

[0071] P w = R T ·(P c - T);

[0072] In the formula: P w is the three-dimensional point in the world three-dimensional coordinate system, and P c is the three-dimensional point in the camera coordinate system; R is the rotation matrix R (a 3×3 orthogonal matrix); T is the translation vector (a 3×1 vector); R T is the inverse of the rotation matrix.

[0073] Figure 2 Shows an example case diagram of converting the image coordinates into global three-dimensional coordinates.

[0074] The GNSS timestamp associated storage includes hardware synchronization and time synchronization. Hardware synchronization is ensured through the NTP protocol or hardware trigger signal, and the clock deviation between all cameras and the GNSS receiver is <1 ms. The video frame capture moment is accurately recorded as tcam; time synchronization is to synchronize the three-dimensional coordinate P wStored as structured data together with the corresponding timestamp tcam:

[0075]

[0076] S4. Based on the trajectory data, use deep learning algorithms to analyze driving behavior characteristics in complex traffic scenarios, including lane-changing decisions, lane-keeping, and following behavior; in this step,

[0077] The lane-changing decision analysis includes the following process:

[0078] Data collection: With the help of on-vehicle cameras, millimeter-wave radars, lidars and other devices, collect data on the vehicle's surrounding environment, such as vehicle position, speed, lane line information, etc., and at the same time record the driver's lane-changing operations.

[0079] Feature extraction: Use a convolutional neural network (CNN) to extract visual features of lane lines and surrounding vehicles from image data; use a recurrent neural network (RNN) or a long short-term memory network (LSTM) to process time series data, such as vehicle speed changes, accelerations, etc., to capture the dynamic characteristics of driving behavior.

[0080] Model construction: Construct a classification model based on deep learning, such as a deep neural network (DNN), a convolutional recurrent neural network (CRNN), etc. Use the extracted features as inputs and output the probability or decision category of the driver's lane change (such as changing lanes to the left, changing lanes to the right, not changing lanes).

[0081] Model training: Use the labeled dataset to train the model, and adopt a cross-entropy loss function to minimize the error between the prediction result and the true label. Update the model parameters through the backpropagation algorithm to improve the accuracy and generalization ability of the model.

[0082] The lane-keeping analysis includes the following process:

[0083] Data collection: Also rely on on-vehicle sensors to collect driving data of the vehicle in the lane, including information such as the vehicle's position and angle relative to the lane line.

[0084] Feature extraction: Use CNN to extract the features of lane lines and determine the boundaries and centerlines of the lanes. Calculate the offset and angle deviation of the vehicle from the lane centerline as the features for lane keeping.

[0085] Model construction: Construct a regression model, such as a multi-layer perceptron (MLP) or a convolutional neural network regression model. Use the extracted features as inputs and output the vehicle's control instructions (such as steering wheel angle, vehicle speed adjustment) to keep the vehicle driving in the lane.

[0086] Model training: Use the supervised learning method to train the model with the actual control instructions of the vehicle as labels. Adopt the mean squared error loss function to measure the difference between the prediction result and the true label, and continuously adjust the model parameters through the optimization algorithm.

[0087] The following is the process of following vehicle behavior analysis:

[0088] Data collection: Collect data such as the relative distance, relative speed, and acceleration between the host vehicle and the preceding vehicle, as well as the motion state information of the preceding vehicle.

[0089] Feature extraction: Extract features such as the rate of change of distance and speed difference between the two vehicles to describe the dynamic characteristics of following vehicle behavior. At the same time, consider the complexity of the traffic scene, such as the degree of road congestion, sudden acceleration and deceleration of the preceding vehicle, and other factors.

[0090] Model construction: Construct a model based on reinforcement learning, such as a deep Q-network (DQN) or a policy gradient algorithm. The goal of the model is to learn the optimal following vehicle strategy to achieve safe and efficient following vehicle driving.

[0091] Model training: Train the reinforcement learning model in a simulated environment, and guide the model to learn optimal behavior through the reward mechanism. The reward function can be designed according to factors such as following vehicle distance, speed stability, and safety, so that the model can make reasonable following vehicle decisions in different traffic scenarios.

[0092] In comprehensive application, integrate the models of lane change decision-making, lane keeping, and following vehicle behavior analysis to construct a complete driving behavior analysis system. This system can monitor the driving state of the vehicle in real time, and provide auxiliary decision-making suggestions for the driver according to different traffic scenarios and driving behavior characteristics, or directly apply them to the control strategy of autonomous vehicles to improve driving safety and comfort.

[0093] In this embodiment, for lane change decision-making, its acquisition includes the following steps:

[0094] S4.1 Calculate the rate of change of the distance between the vehicle and the adjacent lane markings, and set a threshold to determine the starting point of lane change; in this step, first use a deep learning model (such as DeepLabv3+) to segment the lane line pixels, convert them into three-dimensional coordinates in the world coordinate system, and perform quadratic polynomial fitting on the three-dimensional coordinate points of the lane line:

[0095] y = ax 2 + bx + c;

[0096] In the formula: (x, y) is the lateral position in the world coordinate system.

[0097] Then obtain the vehicle centroid coordinates (x υ , y υ) Calculate the heading angle based on the differential of the vehicle trajectory:

[0098]

[0099] Where: Δx = x t - x t-1 and Δy = y t - y t-1 ;

[0100] Project the vehicle's centroid (x υ , y υ ) onto the lane line model and calculate the vertical distance d:

[0101]

[0102] Calculate the rate of change:

[0103]

[0104] In the formula: Δt is the video frame interval;

[0105] Use a 3-frame moving average to eliminate noise:

[0106]

[0107] In this embodiment, the threshold is set according to the empirical threshold method. By statistically analyzing the range during lane change through experiments, the threshold δ th is set to 0.3 - 0.5 m / s. When three consecutive frames satisfy and the lateral acceleration is greater than a th (0.5 m / s 2 ), it is determined that the starting point of the lane change begins.

[0108] Illustrative example:

[0109] Suppose the lateral distance change of a vehicle within 3 seconds is as follows:

[0110]

[0111] In the above table, if δ th = 0.4 m / s, when t = 0.099 s, the absolute value exceeds the threshold, and the lateral acceleration a = -68.18 m / s, it is determined that the starting point of the lane change begins.

[0112] S4.2 Analyze the relative speed and distance of surrounding vehicles based on the Wiedemann psychophysical model to determine the type of lane-changing intention. In this step, the Wiedemann model is mainly used to describe the decision-making process of drivers when changing lanes, taking into account the relative speed and distance of surrounding vehicles, as well as the psychological factors of drivers. The core of the Wiedemann model is to calculate the safe distance and relative speed of the driver. The safe distance usually consists of two parts: the reaction distance at the current vehicle speed and the braking distance. The reaction distance is the distance traveled by the driver during the reaction time, and the braking distance is the distance from the start of braking to the stop of the vehicle. The sum of these two parts constitutes the safe distance. Then, the relative speed refers to the speed difference between the vehicle in the target lane and the own vehicle. If the relative speed is positive, it means the vehicle in the target lane is faster than the own vehicle; otherwise, it is slower. Combining the safe distance and relative speed, the risk and type of lane-changing intention can be judged. For example, if the relative speed of the surrounding vehicles is small and the distance is far, the lane-changing intention of the vehicle may be to overtake; if the relative speed of the surrounding vehicles is large and the distance is close, the lane-changing intention of the vehicle may be to avoid collision or find a more suitable driving lane. Figure 3 Shows a schematic diagram of the distribution of trajectory data according to the Wiedemann model.

[0113] S5. Calibrate the driver model parameters in the traffic flow simulation according to the driving behavior characteristics to generate a driver model; in this step, according to the driving behavior analysis results, combined with the driving behavior datasets of Austria and Zhejiang, calibrate the driver model parameters in the traffic flow simulation in the driver model generation module. For example, adjust parameters such as the aggressiveness of lane-changing, safe distance, and path planning preference to generate driver models with different socio-cultural characteristics. By continuously optimizing the parameter configuration, the generated driver model can accurately simulate the local driving behavior.

[0114] S6. Build a virtual test framework to verify the performance of the autonomous driving function under the calibrated driver model through co-simulation. In this step, in the virtual test and verification module, build a virtual test framework based on GaiA software and PTV Vissim. Conduct static scenario modeling, including road generation and environmental building creation, set various attributes of the road, such as the number of lanes, length, slope, etc., and add various building models. Generate dynamic traffic flow to simulate real traffic conditions. Integrate the multi-body vehicle dynamics model to consider the motion characteristics of the vehicle. Realize the data interaction between GaiA and Vissim through dynamic link libraries for co-simulation. During the simulation process, verify the performance of the autonomous driving function under the calibrated driver model, and at the same time use cloud parallel virtual testing to achieve automated evaluation of a large number of scenarios and generation of test reports through distributed computing. According to the test results, further optimize the autonomous driving system and driver model. Figure 4 To randomly generate driving behaviors and traffic flow simulation effects.

[0115] Furthermore, corresponding tests are conducted on the solution of the present invention:

[0116] I. Experimental data of object detection and tracking:

[0117] YOLOv5 object detection performance:

[0118]

[0119]

[0120] JPDA tracking algorithm performance:

[0121] Scene type MOTA IDF1 Number of trajectory breaks Highway 92.3％ 89.7％ 2.1 times per hour Urban intersection 88.5％ 85.2％ 4.3 times per hour Multi-lane lane change area 89.8％ 86.9％ 3.5 times per hour

[0122] II. Experimental data of driving behavior feature extraction:

[0123] Verification of lane change starting point detection:

[0124] Dataset: A2 Expressway in Austria, Hangzhou Ring Expressway in Zhejiang Sample size: 1000 lane change behaviors are collected for each scenario;

[0125] Detection accuracy:

[0126] Scene type Average starting point detection error (ms) Accuracy rate Highway entrance 125 ms 94.7％ Urban intersection 158 ms 91.3％ Multi-lane lane change area 112 ms 95.8％

[0127] Verification of Wiedemann model parameters:

[0128] Comparison of key parameters (Austria vs Zhejiang):

[0129]

[0130]

[0131] III. Experimental data of driver model calibration:

[0132] Parameter sensitivity analysis:

[0133] Influence of lane change aggressiveness parameter (0 - 1) on simulation results:

[0134] Parameter value Average lane change duration (s) Minimum distance between adjacent lanes (m) 0.3 5.8 2.2 0.6 4.2 1.8 0.9 3.1 1.5

[0135] Verification of model authenticity:

[0136] Comparison of virtual simulation and real trajectory:

[0137] Index Austria dataset error Zhejiang dataset error Lateral position deviation (m) 0.12 0.15 Speed fluctuation (km / h) 1.8 2.1 <![CDATA[Standard deviation of acceleration (m / s 2 )]]> 0.25 0.30

[0138] IV. Joint simulation test data:

[0139] Verification of autonomous driving function

[0140] Test results under the calibration model:

[0141]

[0142]

[0143] Cloud parallel test efficiency:

[0144] Distributed computing cluster configuration: 50 nodes, 8 GPUs per node

[0145] Test throughput:

[0146]

[0147] V. Experimental conclusions

[0148] 1. The object detection and tracking module maintains an accuracy of >90% in complex scenarios.

[0149] 2. The detection error of the lane change starting point is controlled within 150 ms, meeting the requirements of driving behavior analysis.

[0150] 3. The calibration of region-differentiated parameters makes the lateral error between the simulation trajectory and the real data <0.2 m.

[0151] 4. The cloud test efficiency is more than 40 times higher than that of traditional methods.

[0152] These experimental data confirm the ability of the method of the present invention to accurately identify the driver's behavior in different traffic scenarios, especially having significant advantages in lane change decision analysis and region-differentiated modeling.

[0153] The above embodiments are only partial implementation manners of the present invention, and can be adjusted and extended according to specific requirements in actual applications. The protection scope of the present invention is not limited to the above embodiments, but also includes various deformations and improvements based on the technical solutions of the present invention.

Claims

1. A driver behavior recognition method for autonomous driving virtual testing and verification, characterized in that: The following steps are involved: S1. Collecting traffic video data of the target area through multi-source sensors, wherein the multi-source sensors include road cameras and GNSS equipment; S2. Classifying dynamic targets in the traffic video data based on a deep neural network and extracting the motion trajectory of the target; S3. Convert the extracted motion trajectory into trajectory data in the global coordinate system and store it in association with the GNSS timestamp; S4. Based on the trajectory data, using a deep learning algorithm to analyze driving behavior characteristics in complex traffic scenarios; S5. Calibrate the driver model parameters in the traffic flow simulation according to the driving behavior characteristics and generate a driver model; S6. Build a virtual test framework to verify the performance of the autonomous driving function under the calibrated driver model through joint simulation.

2. The driver behavior recognition method for autonomous driving virtual testing and verification according to claim 1, characterized in that: In step S2, dynamic target objects are classified into buses, trucks, motorcycles, bicycles and pedestrians, and continuous extraction of trajectories is achieved through target detection algorithms and multi-target tracking algorithms.

3. The driver behavior recognition method for autonomous driving virtual testing and verification according to claim 1, characterized in that: In step S3, the extracted motion trajectory is converted into trajectory data in the global coordinate system and stored in association with the GNSS timestamp. The specific process is: Convert from image coordinate system to camera coordinate system: Where: (u, ν) is the image pixel coordinate, Z is the z-axis coordinate of the target point in the camera coordinate system; the camera intrinsic parameter matrix K is obtained by Zhang Zhengyou calibration method: Where: f x , f y They represent the focal length coordinates (in pixels), (u0, ν0) are the coordinates of the principal point of the image, Convert from the camera coordinate system to the world 3D coordinate system: P w =R T ·(P c -T); Where: P w is a 3D point in the world 3D coordinate system, P c is a three-dimensional point in the camera coordinate system; R is the rotation matrix R; T is the translation vector; R T is the inverse of the rotation matrix; The GNSS timestamp associated storage includes hardware synchronization and time synchronization. Hardware synchronization is ensured by NTP protocol or hardware trigger signal. The clock deviation between all cameras and GNSS receivers is less than 1ms. The video frame capture time is accurately recorded as tcam. Time synchronization is to convert the three-dimensional coordinate P w The corresponding timestamp tcam is stored as structured data:

4. The driver behavior recognition method for autonomous driving virtual testing and verification according to claim 1, characterized in that: In step S4, complex traffic scenarios include highway entrances, exits, and multi-lane intersections, and the driving behavior characteristics include lane change decisions, lane keeping, and vehicle following behaviors.

5. The driver behavior recognition method for autonomous driving virtual testing and verification according to claim 4, characterized in that: The lane change decision acquisition comprises the following steps: S4.1 calculates the rate of change of the distance between the vehicle and the adjacent lane marking and sets a threshold to determine the starting point of lane change; S4.2 analyzes the relative speed and distance of surrounding vehicles based on the Wiedemann psychophysical model to determine the type of lane change intention.

6. The driver behavior recognition method for autonomous driving virtual testing and verification according to claim 1, characterized in that: In step S5, regional differences are calibrated by comparing driving behavior data sets from different regions to calibrate model parameters, including lane change aggressiveness, safety distance, and path planning preference.

7. The driver behavior recognition method for autonomous driving virtual testing and verification according to claim 1, characterized in that: In step S6, the virtual test framework is implemented based on the joint simulation of GaiA software and PTV Vissim, including static scene modeling, dynamic traffic flow generation and multi-body vehicle dynamics model integration.

8. The driver behavior recognition method for autonomous driving virtual testing and verification according to claim 7, characterized in that: The joint simulation realizes data interaction between GaiA and Vissim through a dynamic link library, supporting closed-loop testing of traffic flow simulation and autonomous driving control algorithms.

Images